File size: 6,264 Bytes
495fe55 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 |
#
# For acknowledgement see accompanying ACKNOWLEDGEMENTS file.
# Copyright (C) 2024 Apple Inc. All rights reserved.
#
from typing import Tuple
import torch
import torch.nn as nn
from timm.models.layers import SqueezeExcite
__all__ = ["ReparamLargeKernelConv"]
class ReparamLargeKernelConv(nn.Module):
"""Building Block of RepLKNet
This class defines overparameterized large kernel conv block
introduced in `RepLKNet <https://arxiv.org/abs/2203.06717>`_
Reference: https://github.com/DingXiaoH/RepLKNet-pytorch
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int,
stride: int,
groups: int,
small_kernel: int,
inference_mode: bool = False,
use_se: bool = False,
activation: nn.Module = nn.GELU(),
) -> None:
"""Construct a ReparamLargeKernelConv module.
Args:
in_channels: Number of input channels.
out_channels: Number of output channels.
kernel_size: Kernel size of the large kernel conv branch.
stride: Stride size. Default: 1
groups: Group number. Default: 1
small_kernel: Kernel size of small kernel conv branch.
inference_mode: If True, instantiates model in inference mode. Default: ``False``
activation: Activation module. Default: ``nn.GELU``
"""
super(ReparamLargeKernelConv, self).__init__()
self.stride = stride
self.groups = groups
self.in_channels = in_channels
self.out_channels = out_channels
self.activation = activation
self.kernel_size = kernel_size
self.small_kernel = small_kernel
self.padding = kernel_size // 2
# Check if SE is requested
if use_se:
self.se = SqueezeExcite(out_channels, rd_ratio=0.25)
else:
self.se = nn.Identity()
if inference_mode:
self.lkb_reparam = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=self.padding,
dilation=1,
groups=groups,
bias=True,
)
else:
self.lkb_origin = self._conv_bn(
kernel_size=kernel_size, padding=self.padding
)
if small_kernel is not None:
assert (
small_kernel <= kernel_size
), "The kernel size for re-param cannot be larger than the large kernel!"
self.small_conv = self._conv_bn(
kernel_size=small_kernel, padding=small_kernel // 2
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Apply forward pass."""
if hasattr(self, "lkb_reparam"):
out = self.lkb_reparam(x)
else:
out = self.lkb_origin(x)
if hasattr(self, "small_conv"):
out += self.small_conv(x)
return self.activation(self.se(out))
def get_kernel_bias(self) -> Tuple[torch.Tensor, torch.Tensor]:
"""Method to obtain re-parameterized kernel and bias.
Reference: https://github.com/DingXiaoH/RepLKNet-pytorch
Returns:
Tuple of (kernel, bias) after fusing branches.
"""
eq_k, eq_b = self._fuse_bn(self.lkb_origin.conv, self.lkb_origin.bn)
if hasattr(self, "small_conv"):
small_k, small_b = self._fuse_bn(self.small_conv.conv, self.small_conv.bn)
eq_b += small_b
eq_k += nn.functional.pad(
small_k, [(self.kernel_size - self.small_kernel) // 2] * 4
)
return eq_k, eq_b
def reparameterize(self) -> None:
"""
Following works like `RepVGG: Making VGG-style ConvNets Great Again` -
https://arxiv.org/pdf/2101.03697.pdf. We re-parameterize multi-branched
architecture used at training time to obtain a plain CNN-like structure
for inference.
"""
eq_k, eq_b = self.get_kernel_bias()
self.lkb_reparam = nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=self.padding,
dilation=self.lkb_origin.conv.dilation,
groups=self.groups,
bias=True,
)
self.lkb_reparam.weight.data = eq_k
self.lkb_reparam.bias.data = eq_b
self.__delattr__("lkb_origin")
if hasattr(self, "small_conv"):
self.__delattr__("small_conv")
@staticmethod
def _fuse_bn(
conv: torch.Tensor, bn: nn.BatchNorm2d
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Method to fuse batchnorm layer with conv layer.
Args:
conv: Convolutional kernel weights.
bn: Batchnorm 2d layer.
Returns:
Tuple of (kernel, bias) after fusing batchnorm.
"""
kernel = conv.weight
running_mean = bn.running_mean
running_var = bn.running_var
gamma = bn.weight
beta = bn.bias
eps = bn.eps
std = (running_var + eps).sqrt()
t = (gamma / std).reshape(-1, 1, 1, 1)
return kernel * t, beta - running_mean * gamma / std
def _conv_bn(self, kernel_size: int, padding: int = 0) -> nn.Sequential:
"""Helper method to construct conv-batchnorm layers.
Args:
kernel_size: Size of the convolution kernel.
padding: Zero-padding size.
Returns:
A nn.Sequential Conv-BN module.
"""
mod_list = nn.Sequential()
mod_list.add_module(
"conv",
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=kernel_size,
stride=self.stride,
padding=padding,
groups=self.groups,
bias=False,
),
)
mod_list.add_module("bn", nn.BatchNorm2d(num_features=self.out_channels))
return mod_list
|